Florida State's Michael Gubanov won a five-year NSF CAREER Award to build an AI model that understands numeric data. He is applying it to cancer research with Moffitt Cancer Center.
A Florida State University researcher has won an NSF CAREER Award to fix a basic weakness in AI: numbers.
Michael Gubanov is an associate professor of computer science. He will build a large language model that understands properties of numeric data, such as number order.
The award provides five years of funding for scientific goals, student training and outreach.
The National Science Foundation calls its Faculty Early Career Development Program its most significant award for early career faculty.
Why Do AI Models Struggle With Numbers?
Large language models answer numeric questions through statistical pattern recognition. They encode numbers as tokens, so they often guess the most probable answer instead of calculating it.
Gubanov gives a simple case. A model that answers four to two plus two does so because four is the likely reply, not because it added.
“Models don’t fully comprehend the order and magnitude of numeric data,” Gubanov said.
He says that limit blocks adoption in medicine. “Currently, LLMs hallucinate and are inherently non-deterministic,” he said.
The result is accuracy barriers where numeric precision matters most. Gubanov plans a model that does not rely on guesswork.
Gubanov states the goal plainly. “Our goal is to develop a model that can instead apply more meaningful quantitative reasoning to better comprehend numerical data.”
FSU says stronger numeric analysis could make AI in clinical practice more trustworthy.
He says numeric data is “often more complex and structured than natural language text.”
FSU says language models often give inaccurate answers to numerical operations for the same reason.
What Will the Model Do in Cancer Research?
The project applies the model to large-scale healthcare data. Gubanov is working with Moffitt Cancer Center in Tampa to test it.
The team will study personalized recommendations for treatments such as chemotherapy and immunotherapy. The model could gauge candidate cancer treatments using factors like a patient’s prior treatments.
“While AI can’t diagnose or treat patients, this advanced AI model is expected to help healthcare professionals make more accurate decisions,” Gubanov said.
FSU says the model could also serve industries that need precise numeric calculations. It names medicine, government, finance and science.
Who Else Backs the Work?
Gubanov founded and supervises BigLab! at FSU. The lab bridges AI and data science with cancer research to improve cancer care.
BigLab! has also received support from the Casey DeSantis Cancer Research Programs’ Florida Cancer Innovation Fund and Amazon Web Services.
Sam Huckaba, dean of FSU’s College of Arts and Sciences, called the award “a mark of excellence.”
He said Gubanov has used AI techniques in novel ways to interpret large-scale health data. “There’s a level of complexity that must be overcome to provide swift and meaningful analysis of these rich datasets.”
Gubanov earned his doctorate in computer science from the University of Washington in 2010. He did postdoctoral research at MIT and joined FSU in 2018.
The project also funds new teaching materials, student research opportunities at BigLab! and outreach on data reasoning.
How Does This Connect to South Florida Research?
Reliability is a live research topic in South Florida. MAIN has covered University of Miami engineers teaching AI to admit uncertainty.
Cancer care is another local thread. MAIN has covered AI entering the tumor board at Miami Cancer Institute.
MAIN has also covered an FSU professor who won a Microsoft grant to study agentic hiring. Florida's universities are drawing AI research funding from both public and private sources.
Gubanov’s project is a five-year effort, so results are a long way off. The accuracy problem it targets is one that finance and healthcare leaders already face.
Sources:
- "FSU data scientist receives NSF CAREER Award to develop AI model to analyze complex numerical data at scale," Florida State University News
- MAIN: UM engineers teach AI to say I'm not sure
- MAIN: AI enters the tumor board, Miami Cancer Institute is in
- MAIN: An FSU professor got a Microsoft grant to study agentic hiring